LOCOGEN: Locality-Guided Approach Solves the Regulatory DNA Design Challenge

Regulatory DNA sequence design is one of the core challenges in synthetic biology and gene therapy, aiming to intervene in disease by precisely controlling gene expression. In recent years, generative AI methods such as diffusion models have been introduced into this field, generating highly active sequences through reward fine-tuning or guidance at sampling time. However, a recent study points out that single-reward evaluation may mask reward hacking, causing generated sequences to appear highly active while actually lacking novelty or plausibility. To address this, the researchers propose LOCOGEN, which reframes the guidance problem as a locality problem, achieving a balance among performance, novelty, and plausibility across multiple cell lines.

The Blind Spot of Existing Methods: Reward Hacking and Evaluation Bias

Current mainstream methods treat regulatory DNA design as an attribute-conditioned generation task: first train a diffusion model to learn the natural sequence distribution, then push generation toward high predicted activity through reward fine-tuning or guidance at sampling time. This paradigm has succeeded in areas such as image generation, but problems arise when it is directly transferred to DNA sequences. The researchers found that when evaluated with only a single reward (such as predicted activity), the model may learn to "cheat"—for example, generating sequences highly similar to the training set but with inflated activity, or producing structurally implausible sequences that fool the predictor. This reward hacking behavior is particularly dangerous in biological sequence design because activity prediction itself is noisy and experimental validation is costly.

To reveal this blind spot, the researchers decomposed the evaluation space into three dimensions: property (i.e., predicted activity), novelty (difference from the training set), and plausibility (whether the sequence resembles natural DNA). Through systematic analysis, they identified an empirically underexplored region: within this region, all three objectives are simultaneously satisfied, yet none of the existing baseline methods succeed. This finding suggests that previous methods focused only on the "magnitude" of guidance while ignoring the "locality" of guidance.

Reward Hacking is a classic failure mode in reinforcement learning and generative model alignment: when a model is optimized to maximize some proxy reward, it may find loopholes in the reward function and produce high-scoring results that do not actually meet the design intent. In image generation, this manifests as image distortion caused by over-optimizing the reward; in biological sequence design it is more insidious—the model can generate highly active sequences nearly identical to the training set (seemingly successful but actually lacking novelty), or generate structurally abnormal sequences that fool the predictor. Because the activity predictor itself is trained on limited experimental data, it has systematic biases and noise, and the model can easily exploit these blind spots. Therefore, judging generation quality by a single reward metric alone is essentially substituting a flawed proxy metric for the true objective.

The Locality Perspective: The Sequence Grammar Space Is Non-Uniform

Why do existing methods fail? The researchers propose that the sequence grammar of regulatory DNA is not uniformly distributed in space. Specifically, the positions of motifs in DNA sequences are highly sensitive to perturbation—changing a single base can disrupt transcription factor binding—whereas background regions allow more exploration, and moderate variation does not significantly affect function. This spatial non-uniformity means that a uniform guidance strength creates a contradiction between protecting motifs and exploring background: too strong guidance destroys motifs, while too weak guidance fails to effectively explore the background.

Therefore, the researchers reformulate guidance as a locality problem: protective constraints should be imposed on motif positions, while exploratory guidance should be applied to background positions. This idea is consistent with natural regulatory logic—functional elements need to be conserved, while spacer regions can tolerate variation. Based on this, they propose the LOCOGEN method.

Motif refers to short fragments in DNA sequences with specific base arrangement patterns (typically 6-20 bp), which are the minimal functional units recognized and bound by regulatory proteins such as transcription factors. If a key base within a motif mutates, it may directly disrupt transcription factor binding and cause substantial changes in gene expression; whereas the spacer sequences between motifs (background regions) often tolerate a higher degree of variation. This characteristic of "conserved functional sites, plastic spacers" is precisely the molecular basis of the spatial non-uniformity of regulatory DNA sequence grammar. Understanding this structural property of motifs is key to grasping why LOCOGEN applies differentiated guidance to different positions—protecting motifs is equivalent to protecting the functional core of the sequence, while opening up the background is equivalent to preserving room for exploration and novelty.

LOCOGEN: Invariant Motif Masking and OOD Damping

The core of LOCOGEN is localized guidance of a frozen diffusion backbone at sampling time, without retraining the backbone. It contains two key components:

• Invariant motif mask learning: Through sequence-level invariance learning, it identifies motif positions that are critical to function. Specifically, the model learns a mask such that at these positions, perturbation of the sequence causes significant changes in predicted activity, thereby protecting them from excessive guidance.
• OOD damping: At background positions, out-of-distribution (OOD) damping is applied, i.e., penalizing generations that deviate too far from the natural distribution, but only in background regions to avoid suppressing novelty.

This approach is performed at sampling time and requires no fine-tuning of the diffusion model, so it has low computational cost and is easy to combine with existing backbones. The authors conducted experiments on three cell lines—HepG2, K562, and SK-N-SH—and the results show that LOCOGEN can reach previously unexplored regions, i.e., improving novelty and plausibility while maintaining high predicted activity.

Out-of-Distribution (OOD) damping draws on the idea of controlling distribution shift in generative models. If guidance is too strong during diffusion model sampling, generated samples may deviate from the natural data manifold and fall into regions the predictor has not been sufficiently trained on, where high predicted activity is unreliable. OOD damping applies a penalty to samples that deviate too far from the natural distribution, pulling generation back to trustworthy regions. LOCOGEN's innovation lies in localizing this penalty: OOD constraints are applied only to background regions and not to motif regions, thereby suppressing implausible sequences without sacrificing functional elements. This "differentiated treatment by region" approach decouples guidance strength from guidance position and avoids the drawback of global damping suppressing novelty.

Multi-Source Validation and Experimental Insights

According to the paper, LOCOGEN achieves a balance among performance, novelty, and plausibility on all three cell lines, while baseline methods cannot satisfy all three simultaneously. This result is cross-validated across multiple sources: data from different cell lines point to a consistent conclusion—that locality-guided approach is an effective way to solve the reward hacking problem. However, the researchers also note that current evaluation still relies on computational predictors, and experimental validation will be future work. In addition, the code is open-sourced on GitHub (https://github.com/Lotfollahi-lab/LocoGen), facilitating community reproduction and extension.

From a broader perspective, this study reminds us that the application of generative AI in scientific design requires more refined evaluation frameworks. Single-metric optimization easily leads to unintended behavior, whereas decomposing the problem into multiple dimensions and considering spatial heterogeneity may provide lessons for other sequence design tasks (such as protein engineering). LOCOGEN's locality approach may become a new starting point for regulatory DNA design.

Guidance at Sampling Time and Reward Fine-tuning are two mainstream methods for controlling generative models. The former modifies the denoising trajectory during inference (e.g., through classifier guidance or classifier-free guidance) without updating model parameters; the latter requires additional training of the model with a reward signal. The advantage of guidance at sampling time is its low computational overhead, flexible switching of rewards, and preservation of the natural sequence distribution learned during pretraining; the disadvantage is that guidance strength is difficult to control precisely and tends to globally distort generation results. LOCOGEN chooses localized guidance at sampling time precisely because of its lightweight nature and composability—a frozen backbone means it can be directly layered onto existing diffusion models without retraining for each new objective.

Future Outlook

Although LOCOGEN performs excellently in computational evaluation, validation through actual biological experiments remains a key step. In addition, how to extend locality-guided approach to more complex regulatory elements (such as enhancer-promoter interactions) is worth exploring. As diffusion models become more widespread in biological sequence design, refined guidance strategies like LOCOGEN are expected to become standard tools, advancing the development of precision gene regulation therapies.

One-line Summary

LOCOGEN solves the reward hacking problem in regulatory DNA design through locality-guided approach, balancing activity, novelty, and plausibility.

Paragraph Summary

Regulatory DNA sequence design aims to precisely control gene expression to intervene in disease. Generative methods such as diffusion models generate highly active sequences through reward fine-tuning or sampling guidance, but single-reward evaluation may mask reward hacking, causing sequences to appear highly active while lacking novelty or plausibility. The researchers propose LOCOGEN, which reframes guidance as a locality problem: using invariant motif masking to protect functionally critical sites and applying OOD damping in background regions to suppress implausible generation, thereby achieving localized guidance of a frozen diffusion model at sampling time. On three cell lines—HepG2, K562, and SK-N-SH—LOCOGEN is the first to simultaneously achieve a balance among property, novelty, and plausibility, outperforming existing baselines. This work emphasizes the importance of multi-dimensional evaluation and spatial heterogeneity modeling for generative scientific design.